AI · LLMs · NLP — MSc Data & AI, Institut Polytechnique de Paris
I build AI systems, enjoy tackling hard problems, and want to make AI more reliable and trustworthy. I still believe I might change the world through what I do. Maybe that's naive. I'm not ready to give it up.
In my research on YAGO and Wikidata I worked on keeping a knowledge base of millions of facts logically consistent: the kind of reliable, structured knowledge that language models still struggle with.
LLMs · 2026
LLM prompting (zero-shot, few-shot, chain-of-thought) vs. fine-tuned BERT vs. TF-IDF on CLINC150. Best LLM strategy: 0.839 macro-F1 — but BERT matches its accuracy at a fraction of the inference cost.
Mistral API · BERT · prompt engineering
Read more → Report ↗
Deep learning · 2025
Is a tweet written by an influencer or an observer? A multitask network with cross-attention between 23 engineered user features and CamemBERT tweet embeddings: 0.83 F1 and 0.92 ROC-AUC in 5-fold cross-validation.
PyTorch · CamemBERT · Optuna
Read more →
Bachelor thesis · 2025
Detecting misogyny in Italian tweets (AMI EVALITA 2020): five tuned classical models against a CNN on GloVe embeddings. The CNN wins with 0.83 macro-F1 — but a well-tuned random forest is only 0.001 behind.
Python · scikit-learn · Keras · GloVe
Read more → Thesis (EN) ↗Thesis (IT) ↗
Complex systems · 2025
Ant foraging simulation where homing emerges from a diffusive "colony pheromone" instead of global coordinates.
Agent-based modelling · Python
Read more → Report ↗
Parallel programming · 2025
Renders scenes of coloured spheres, parallelized with OpenMP and writing PPM images through mmap.
C · OpenMP
Read more →
Software engineering · 2025
The game was the excuse: the goal was learning how a team ships software. Test-driven development with 70 JUnit and Mockito tests, continuous integration on GitHub Actions building and testing every push, and steady refactoring.
Java · Gradle · JUnit · GitHub Actions
Databases · 2024
A relational database for optimising Milan’s transport network, designed end to end: requirements, E-R schema, restructuring and normalisation, then SQL tables, operations and triggers, queried from Python.
SQL · E-R modelling · Python
Read more → Report ↗
Programming · 2025
Assembler and interpreter for the Little Man Computer, with step-by-step execution and error handling.
Python
Read more →
An empirical analysis showing that punning — entities acting both as classes and as individuals — affects thousands of Wikidata classes, a formalization of the requirements for representing it, and a comparison of knowledge representation formalisms.
A new version of YAGO that, for the first time, quantifies and shows not only the Wikidata statements kept in the knowledge base but also those filtered out — shedding light on the inconsistencies of large collaborative knowledge bases and the trade-offs of enforcing a schema.
YAGO Knowledge Base Research, Knowledge Representation
Research on keeping a knowledge base built from Wikidata logically consistent, leading to two ISWC 2026 papers.
Studio Ingegneria Web Web & Database Developer
Built management software and databases, leading small projects from requirements to delivery.
Member, IASEAI
MSc in Data and Artificial Intelligence
Institut Polytechnique de Paris · Palaiseau, France
Large-scale Generative Models for NLP & Speech, Advanced Deep Learning, Text Mining & NLP, Reinforcement Learning, Graph Learning, Verification of Neural Networks, Optimisation & Responsible AI
BSc in Artificial Intelligence and Data Analytics
Università degli Studi di Trieste · Trieste, Italy
Fencer since 2015, former scout leader, and volunteer teacher in Nepal (2024) and builder in Peru (2021). Hitachi merit scholarship holder. I enjoy Japanese literature, sailing, hiking and reading about international politics.